SURFNet: A Surface-Aware UAV–Satellite Geolocation Framework via Feature Aggregation and Dual Positional Encoding

Cross-view geo-localization (CVGL) between UAV and satellite imagery is crucial for GNSS-denied navigation but brittle under domain shift due to background shortcuts and layout misalignment. We present SURFNet, a surface-aware two-stage framework that boosts accuracy and cross-dataset generalization via explicit ground-structure modeling. A Satellite Semantic Augmentation Algorithm (SSAA) preserves salient surface elements and synthesizes structure-preserving pseudo-backgrounds to diversify satellite training data. A Dual Positional Encoding (DPE) combines learnable 2D embeddings with rotation-aware encodings, while a Ground-Aware Aggregation Block (GAB) performs dynamic multi-branch integration of local–global cues.On University-1652 and SUES-200, SURFNet delivers competitive in-dataset results and achieves state-of-the-art zero-shot generalization from University-1652→SUES-200 across altitudes. Additional transfers underline deployability: University-1652→DenseUAV exposes a large style/coverage gap that is substantially closed by lightweight fine-tuning (freezing the backbone, adapting small heads on a 50-class subset), while SUES-200→University-1652 shows expected asymmetry tied to the smaller source diversity of SUES-200. Subset studies on SUES-200 (dense/sparse buildings, water scenes, natural landscapes) further indicate that SURFNet exploits general surface saliency—not only the four SSAA categories—to generalize across scene types. Overall, surface-aware augmentation plus structure-guided aggregation yields features that are both accurate and adaptable for real-world CVGL.